Skip to content

.prompt.md

Latest commit

 

History

History
53 lines (43 loc) · 3.26 KB

.prompt.md

File metadata and controls

53 lines (43 loc) · 3.26 KB
Raw
title Commit Prompt Template
description We will generate a commit to main, or if requested, to a feature branch. The agent should detect which files it changed, stage only those files, create a concise commit message, and optionally push the branch. The agent must report the exact git commands used and the resulting commit hash and message.
agent ask
model GPT-5 mini (copilot)

commit prompt

Purpose: Instruct a coding agent how to commit the files it changed.

Inputs (use when invoking):

  • task: short title describing the change (one line)
  • filesChanged: optional list of file paths the agent modified (if known)
  • branchPreference: either main (default) or a feature branch name
  • commitMessage: optional commit message; if not provided, agent should synthesize one
  • push: boolean (false by default) — whether to push to the remote after commit

Behavior / Steps the agent MUST follow:

  1. Detect what files were changed by the agent. Prefer git status --porcelain or the provided filesChanged list. Only stage and commit those files.
  2. If branchPreference is a feature branch name (not main), create and switch to that branch (e.g., git checkout -b <branch>), then commit there. If branchPreference is main or omitted, commit on main.
  3. Stage only the changed files: git add -- <files...> (do not add unrelated files).
  4. Create a single commit using the provided commitMessage or synthesize a concise message following this pattern: <area>: <short description> (e.g., dashboard: prevent widget drag when layout locked). Use present-tense verbs and keep it to ~50 chars for the subject.
  5. If push is true, push the branch to the remote: git push origin <branch>.
  6. Return a concise report showing: branch name, commit short hash, committed files list, the exact git commands run, and suggested next steps (e.g., open PR or push).

Constraints / Safety:

  • Do not stage or commit files the agent did not intentionally change.
  • Avoid force-pushing or rewriting published history.
  • If any staged file appears surprising, pause and ask the user before committing.

Output format (agent reply after committing):

  • Branch: <branch>
  • Commit: <short-hash> <commit message>
  • Files committed: list of paths
  • Commands run: show the exact git commands used
  • Next steps: one-line suggestion (e.g., git push origin <branch> or Create PR from <branch>)

Examples:

  • Commit to main (default):

    • Inputs: { task: "fix dashboard drag", branchPreference: "main", commitMessage: "dashboard: lock layout drag" }
    • Behavior: stage modified files, commit on main, report commit hash, do not push unless push:true.

  • Commit to feature branch:

    • Inputs: { task: "dashboard-lock-fix", branchPreference: "feature/dashboard-lock", commitMessage: "dashboard: prevent drag when locked", push: true }
    • Behavior: create feature/dashboard-lock, stage files, commit, push branch, return push command and PR suggestion.

Iteration guidance:

  • If the agent cannot run git (permissions or repo not present), it should produce the exact sequence of commands to run locally and the patch (apply_patch) for review.
  • When uncertain which files to include, ask a single clarifying question listing the candidate files.

Where to save: repository root as .prompt.md (this file).